When do I conclude that my NLMIXED model is not converging?

[marcioadg]

Actually it is converging, FCONV, but the MaxGrad is not below 0.0001 or even 0.01. The parameter estimates are exactly the same as my initial values on the Parms statement. I am pretty sure I have parameters close enough. Is there anyway around it? 

When I run the linear -plateau model it is ok. The problem is when I run the curvilinear-plateau model.

Thank you

Marcio

[SteveDenham · Accepted solution]

Try different starting values.  It may be that you are in a local extremum.  Grid search may be the best way to go after this.  Also, check out the algorithm you are using for optimization.  This means examining alternate optimization techniques, or altering the update method or linesearch method if you are using an optimization techniques that allows alternatives in these areas.

Steve Denham

[marcioadg]

I will try that, but can I compare IC from between models using different optimization techniques? Or than I would need to chance the opt tech for all my models?

Currently I am using the default for NLMIXED (Dual Quasi-Newton).

[SteveDenham]

I would say it's probably a good idea to compare IC's based on the same optimization technique.  If you have trouble with one model/dataset, changing to a different method to get it to work for that instance means you should, at the very least, test that method on the other models/datasets.  It would be optimal to use the same technique on all candidate models.

Steve Denham